The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in several scripts (pipeline_runner.py, compile_latex_with_bibtex.py, convert_latex_to_word.py) to execute LaTeX compilation (xelatex, bibtex) and document conversion (pandoc). This is the primary purpose of the skill. To mitigate risks, the author implemented a security module scripts/path_scope.py which enforces that all file operations and command arguments are restricted to the designated working directory.
[EXTERNAL_DOWNLOADS]: The skill fetches bibliographic metadata and abstracts from established academic services including api.openalex.org, api.semanticscholar.org, and api.crossref.org. These are well-known, trusted academic infrastructure providers. The data retrieval is performed via the requests and urllib libraries in a controlled manner.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests untrusted paper titles and abstracts from external APIs into LLM prompts for relevance scoring (references/ai_scoring_prompt.md).
Ingestion points: External paper metadata is fetched in scripts/openalex_search.py, scripts/semantic_scholar_search.py, and scripts/crossref_search.py and stored in JSONL files.
Boundary markers: The scoring prompt in references/ai_scoring_prompt.md uses clear JSON-style field labels (e.g., "标题: {title}", "摘要: {abstract}") to separate untrusted data from instructions.
Capability inventory: The skill can execute shell commands via pipeline_runner.py to compile documents.
Sanitization: The skill includes _sanitize_unicode in scripts/build_reference_bib_from_papers.py to remove Unicode control characters that could interfere with LaTeX compilation or hide malicious content.

research-literature-review